Automatic Indexing and Generating of Content Graphs from Unrestricted Text

F or qu ite som e tim e, I have been exp lorin g the surface signals o f language, and try in g to put them to as m uch use as possib le , prim arily in m orp h ology -based p art-o f-sp eech assignm ent (K ällgren 1984a,b ,c, 1985) and pattern -based syn tac­ tic analysis (K ä llgren 1987). T h is kind o f large-scale, probabilistic parsing on the basis o f m orph olog ica l and syn tactic patterns has lately com e to use in several p ro je cts . S om e m odels that have been docu m en ted axe the U C R E L parser in E ngland for the B row n and L O B co rp o ra (G arside & Leech 1987), the V O L S U N G A parser in the U S A for the B row n corpu s (D eR ose 1988), K en C hu rch ’s stoch astic m odels (C h urch 1988), as well as oth er w ork in the US (B lack 1988), bu t I am sure w ork a lon g these lines is go in g on in several places. T h e im petu s beh in d the w orks ju st m entioned is m ainly a need for analyzing large am ou n ts o f unrestricted text in a w ay that is n ot t o o resource-dem anding, e ither on tim e or on com p u tin g pow er. A s a secon dary goa l, I have seen the needs o f large-sca le in form ation retrieval. K eep in g m y original surface-orientation , I have gon e further from the analysis in to parts-o f-speech and constituents and started to lo o k at the ex traction and representation o f som e kind o f ‘ con ten t’ from the surface o f texts , w ith ou t any kind o f know ledge base support. T h is m ight seem qu ite im possib le . M an y o f the o th er papers in this volum e deal w ith h ow unavailable in ferences are to the com prehension o f text, and o f course they are right. I f th e aim is t o bu ild a com pu terized system that will in any way sim ulate language understanding, it is necessary to have a large know ledge base and m echanism s for m aking inferences from it, but there are also applications w here the hum an know ledge and in ferencing capacities can be used instead. M y ap p roach in the experim en t to b e reported here has been to let each on e do